Stanford-Led Work Shows Self-Organizing Agent Teams Outperform Best Single Model

Researchers from Stanford, Together AI, Emory and other institutions released a new paper, "Self-Organizing Agent Teams Learn to Reason Together," proposing Self-Organizing Agent Teams (SAT): teams of fixed members that, instead of relying on hand-designed collaboration protocols, learn reusable role assignments, dialogue phases, participation rules and information-sharing strategies from past collaboration experience. The work was partly motivated by recent debates on multi-agent collaboration sparked by Navier-Stokes solver progress. Team member jamesyzou repeatedly promoted the paper, and Wisconsin researcher Dimitris Papailiopoulos shared it as a representative work on test-time collaboration.

Confirmed

The team consists of three models — o3-mini, Claude Sonnet 4 and DeepSeek-V3 — averaging 66.7% across five math and physics benchmarks.
The strongest single model scores 48.8% under the same setting, so the team clearly outperforms its best member; per one share (m4), it also beats ideal routing.
The method lets the team learn collaboration strategies itself (roles, dialogue phases, participation rules, information sharing) rather than using preset protocols.
Authors include Aneesh Pappu and others, affiliated with Stanford, Together AI, Emory, etc.

Why it matters

The results provide quantitative evidence that multi-agent collaboration can deliver reasoning gains beyond a single model, and that collaboration strategies themselves can be learned and reused.
Released by jamesyzou's group and endorsed by multiple researchers, it is a notable new work in test-time multi-agent collaboration with reference value for future agent system design.

2026-09-23 ~ 2026-09-24 · 6 related posts

Primary sources

New research shows self-organizing agent teams learn to reason together, outperforming solo agents — james_y_zou · 2026-09-23 00:06 (UTC+08:00)
Self-organizing agent teams hit 66.7% accuracy, beating strongest single member's 48.8% — airesearch12 · 2026-09-23 19:51 (UTC+08:00)

[source] New research shows self-organizing agent teams learn to reason together, outperforming solo agents — james_y_zou · 2026-09-23
New research: self-organizing agent teams learn to reason together — james_y_zou · 2026-09-23
Self-organizing agent teams learn to reason together: new study maps when communication pays off — james_y_zou · 2026-09-23
[source] Self-organizing agent teams hit 66.7% accuracy, beating strongest single member's 48.8% — airesearch12 · 2026-09-23
Stanford paper: self-organizing agent team hits 66.7%, beats best single model's 48.8% — dair_ai · 2026-09-23

1 near-duplicate retellings: james_y_zou